NFL Week 6 AI Predictions: Daggy, Kalshi and the Bookies Struggle

Every week this NFL season, we're pitting 8 AI agents (4 with Daggy's statistical engine and 4 without) against Kalshi's prediction market. Here's the Week 6 recap.

NFL Week 6 AI Predictions vs Kalshi: Brier Scores

Kalshi and Daggy both had their worst weeks of the season, but Daggy's was slightly less bad, beating Kalshi 0.2669 to 0.2673, although Kalshi picked two more winners. The week was full of upsets: the Rams lost to the Cardinals; the 4-1 Jaguars lost to the 1-4 Texans; the zero-win Titans beat the supposedly less-awful Colts. Similarly, after losing to the then 1-4 Jets last week, the Browns came back and beat the 4-1 Ravens.

All this is to say that the season's been difficult to predict, but our agents and their machine learning engine should make their predictions better.

• This was Daggy and Kalshi's worst week of the season: both Brier scores were worse than if they had predicted every game as 50-50. The closing-line books also performed terribly, with only about 10 percent of weeks doing worse going back to 2006.
• 11 of 14 games were decided by one score (8 points or less). Going back to 2000, only about 1 week in 64 has a share that high. Typically, about half the games fall within 8 points.
• Underdogs won 4 of 6 division games (Arizona, Tennessee, Cleveland, Houston).

How the Agents Forecast

Following Weeks 4 and 5 we continued to let our agents build memory streams. They analyzed each game and created heuristics. For example, Kimi with Daggy's engine noted the following rule:

favourites have gone 12-1, and moving toward the underdog cost Brier … favourites with an Elo gap over 100 and no quarterback change have gone 12-1

It used this information to put an 82 percent probability on the Rams, 72 percent on the Colts and 69% on the Ravens. All those teams lost. Of course, we don't necessarily have enough data to say these were bad decisions, but they cost the agents this week.

Changes for Week 7

This week we're giving the agents access to Daggy's statistical engine when they do recaps. Since the agents already create heuristics, such as the above on favorites with Elo gaps, they should be able to be more rigorous and test these heuristics against the data, which Daggy's toolkit allows them to do.

For instance, our Opus agent had a card that said favorites went 12-1, or that they almost always win, and it used this to not push probability toward an underdog. For example, before TEN@IND it discussed reasons to doubt the Colts, writing: "favourites with an Elo gap over 100 and no quarterback change have gone 12-1 … Together these justify only a tiny trim, which lands at 0.77." But now with access to data analysis tools it made this rule much more refined: "all 2026 favourites won 70% (n=40) vs 72% by Elo; 2000-25 72% (n=3055)." This shows that big favorites win about 70% of the time, not almost always. This should allow it to push probability toward underdogs when other conditions arise. We will see if this improves the models going forward.